Multilingual Speech Transcription System for Kazakh, Russian and English Languages

Abstract

This paper presents a novel multilingual speech transcription system designed for Kazakh, Russian, and English languages. Unlike existing solutions such as OpenAI Whisper and Kaldi-based offline models, the proposed system introduces three key innovations: (1) a specialized preprocessing pipeline optimized for Kazakh phonetic characteristics, (2) dynamic language detection with confidence scoring, and (3) fine-tuned acoustic models trained on a comprehensive trilingual dataset of 450 hours. The system achieves a 23% better Word Error Rate (WER) for Kazakh (8.7% vs. 11.3%) compared to Whisper and a 15% improvement for code-switched utterances. Statistical significance testing using paired t-tests (p < 0.001) and 95% confidence intervals confirm the superiority of the proposed approach across all target languages. Unlike existing solutions such as OpenAI Whisper which achieves 11.3% WER for Kazakh, the proposed system demonstrates 23% improvement with 8.7% WER (p < 0.001, 95% CI: [8.1%, 9.3%]), while maintaining complete offline functionality and specialized optimization for Central Asian linguistic patterns. However, OpenAI Whisper supports 99 languages, while proposed system supports 3 languages, and has better zero‑shot performance.